Normally, training a model means gathering everyone's data in one place, which raises privacy concerns. Federated learning flips this: the model is sent to each device, learns locally from data that never leaves, and only the resulting adjustments are sent back and combined.
This keeps sensitive information where it originated while still improving a shared model. It is more complex to coordinate and the updates themselves need safeguards, but it offers a way to learn from data that could not or should not be pooled.